Generative sequential recommendation model and construction method for recommendation fairness

Abstract

The present invention relates to a generative sequence recommendation model and construction method for recommendation fairness, the method comprising: injecting random noise into user interaction behavior to obtain a damaged interaction representation that obeys a Gaussian distribution, then designing a sequence denoising model in the reverse restoration process, and using a dynamic routing mechanism to mine multiple interest representations in user historical interaction behaviors, and performing two-stage training in combination with user-side modeling to complete denoising; imitating the forward noise addition process to make the project representation noisy and then use it to iteratively reversely reconstruct the generated project. Compared with the traditional sequence recommendation model, the present invention, on the one hand, constructs a generative recommendation paradigm based on a diffusion model, and models uncertainty and alleviates noise by learning the statistical laws of data and generating new data based on it; on the other hand, it avoids the influence of factors such as user sensitive characteristics that may lead to discrimination or unequal treatment on the recommendation effect, and further promotes the diversity of the generative sequence recommendation model by improving the fairness of the recommendation.

Description

Technical Field

[0001] The present invention relates to the technical field of sequence recommendation models, and in particular to a generative sequence recommendation model oriented to recommendation fairness and a construction method thereof. Background Art

[0002] The development of sequence recommendation models has gone through two stages: first, modeling the sequence recommendation task as a Markov decision process; second, with the rapid development of deep learning methods, recursive neural networks, Transformers, convolutional neural networks, and graph neural networks have all been successfully applied to sequence recommendation tasks. The key point of this type of model is to represent items as fixed vectors, and use the similarity of item vectors as the basis for predicting future interactive items. Although this type of model has achieved good results, the fixed vector representation will lead to three problems: first, in the real recommendation scenario, the user's interests and the topic distribution of the project are diverse. At the same time, if the recommendation scenario and period change, the user's interests will also change dynamically, and the singleness and staticness of the fixed vector representation are contrary to the characteristics of users in the real recommendation scenario; second, the anti-noise ability is weak, and the training of fixed item vector representation is easily affected by data bias. Unstable vector representation will further aggravate the bias of the model recommendation results; third, there is a recommendation bias caused by user sensitive features that are easy to cause discrimination (such as gender and age).

[0003] Therefore, it is urgent to design a new technical solution to comprehensively solve the problems existing in the prior art. Summary of the invention

[0004] The purpose of the present invention is to provide a generative sequence recommendation model and construction method for recommendation fairness. Compared with the problems existing in traditional sequence recommendation models, the present invention, on the one hand, constructs a generative recommendation paradigm based on a diffusion model, and models uncertainty and alleviates noise by learning the statistical laws of data and generating new data based on it; on the other hand, it avoids the influence of factors such as user sensitive characteristics that may lead to discrimination or unequal treatment on the recommendation effect, and further promotes the diversity of the generative sequence recommendation model by improving the fairness of recommendations.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for constructing a generative sequence recommendation model for recommendation fairness includes the following steps:

[0007] Training phase: random noise is injected into the user interaction behavior to obtain a damaged interaction representation that follows a Gaussian distribution; then denoising is completed in the reverse restoration process to provide a mapping from the damaged state to the original state;

[0008] Inference phase: imitate the forward denoising process to noisy the first n-1 items and then iteratively generate items in the reverse reconstruction process.

[0009] A further solution is to inject random noise into the user interaction behavior, and obtain the damaged interaction representation that obeys the Gaussian distribution as follows:

[0010] First, the interaction sequence of user u The Transformer sequence encoder is used to encode That is, E = Embedding (S u ), where Embedding∈R |I|*d is an embedding representation vector table with dimension d;

[0011] In the forward diffusion process, the interaction sequence Each item except the first one is embedded in Diffusion forward; and E n Expressed as and represents the initial encoding state of the item interacted by user u at the t∈[2,T]th moment when no diffusion occurs. The forward diffusion formula is as follows:

[0012]

[0013] in, β i Sampling is performed in a linearly increasing manner; ε obeys the standard Gaussian distribution Interaction Sequence Each project in The final code is Represents the noise-added encoding of the items interacted by user u at the t∈[2,T]th moment after f diffusion processes.

[0014] A further solution is to design a sequence denoising model based on Transformer during the reverse restoration process of the training phase, and use a dynamic routing mechanism to mine the multi-interest representations in the user's historical interaction behavior, combined with user-side modeling for two-stage training;

[0015] In the first stage, a regularized auxiliary loss is added to eliminate the bias caused by user-sensitive features in the multi-interest representation, and then the processed multi-interest representation is used to correct the results of the sequence denoising model;

[0016] In the second stage, the user’s remaining features are used to train the user’s personalized tower;

[0017] During the reverse reconstruction process of the inference phase, items are generated iteratively using posterior Bayesian methods based on the trained model.

[0018] A further solution is that the forward process of the sequence denoising model is:

[0019]

[0020] In formula (2), f θ The input of (·) is the sequence code E0 after being affected by the noise target term. The output is the sequence prediction result for the next moment Used for subsequent iterative denoising and restoration;

[0021] Formula (3) is the influence mode of adding noise to the input sequence. represents the forward diffusion and noise-added encoding of the target item at time t+1; emb step Indicates the encoding of the reconstruction step the target item is currently in; t Indicates the project code at time t The degree of influence of the noise-added target item information;

[0022] In formula (4), λ t It integrates the original item information of the sequence and the noise target item information by bitwise multiplication. position Encodes the location of the item being imported.

[0023] A further solution is to use the dynamic routing mechanism to mine the multi-interest representations in the user's historical interaction behavior. Specifically, the target item embedding and capsule interest vector are used to complete the label-aware attention calculation process, and the user sequence encoding As input, the B2I dynamic routing algorithm outputs a set of capsule interest vectors that can represent the user's diverse intentions. After extracting multiple interest capsules, the output of the sequence encoding model is used As the target item, label-aware attention calculation is performed and finally we get Specific reference formula (5):

[0024]

[0025] in, p is the index, and finally we get the multi-interest representation

[0026] A further solution is to use the processed multi-interest representation to correct the result of the sequence denoising model. Specifically, the user personalized tower and regularization loss are used to correct the user multi-interest representation of the target item:

[0027] First, the user code output by the first stage user personalization tower is expressed as Combining multiple user interest representations Complete the inner product regularization operation as shown in formula (6) as the auxiliary loss of the first stage training;

[0028]

[0029] Afterwards, the result of the sequence denoising model is corrected by using the multi-interest weighting method, and the final denoising restoration is obtained through formula (7) and used as the recommended item code

[0030]

[0031] in, In the second stage of training, the user information is added to model personalization, and the user tower features are replaced with user non-sensitive features for model training. Replace with The item code is generated in the form of formula (7); in addition, the inference stage also uses the formula of the second stage training stage to generate the item code.

[0032] In addition, the present invention also provides a model constructed by the above-mentioned method for constructing a generative sequence recommendation model for recommendation fairness.

[0033] Compared with the problems existing in the traditional sequence recommendation model, the present invention has the following beneficial effects:

[0034] 1. Compared with the currently disclosed sequence recommendation modeling work, this paper innovatively designs a generative sequence recommendation model by combining the diffusion model, which is conducive to modeling uncertainty and alleviating noise, thereby improving recommendation accuracy, and provides a promising development direction for the generative recommendation model;

[0035] 2. The present invention focuses on the fairness modeling of the recommendation process, which can solve the "specific group impression" bias caused by user sensitive characteristics in the traditional sequence recommendation model, and thus can improve the diversity of recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of the forward (diffusion) phase of training and the reverse (reconstruction) phase of reasoning in this embodiment;

[0037] Figure 2 This is a structural diagram of the sequence denoising model of this embodiment. DETAILED DESCRIPTION

[0038] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0039] Example 1

[0040] The technical solution adopted by the present invention is as follows Figure 1 and Figure 2 As shown in FIG, the method for constructing a generative sequence recommendation model for recommendation fairness includes the following steps:

[0041] 1. Applying the Diffusion Model to Sequential Recommendation Tasks

[0042] First, the interaction sequence of user u Encoded as

[0043] Right now Where, Embedding∈R |I|*d It is an embedding representation vector table with dimension d.

[0044] In the forward diffusion process, the sequence Each item except the first one is embedded in Perform forward diffusion. Specifically, the diffusion step f∈[1,N] is obtained by sampling, where N represents the maximum diffusion step.

[0045] In order to better define the modeling process, this embodiment will n Expressed as For example, represents the initial encoding state of the item interacted by user u at the t∈[2,T]th moment when no diffusion occurs. The forward diffusion formula is as follows:

[0046]

[0047] in, β i The sampling is performed in a linearly increasing manner. In addition, ε obeys the standard Gaussian distribution sequence Each project in The final code is Represents the noise-added encoding of the items interacted by user u at the t∈[2,T]th moment after f diffusion processes.

[0048] During the training process, this embodiment uses the forward noise embedding E f As input, we pass the sequence denoising model to obtain the predicted target item embedding Then substitute it into the matching loss to complete the training of the sequence denoising model.

[0049] In the reverse reconstruction process, from a noise sequence x f Initially, this embodiment restores the initial project code x0 by iterative denoising. Specifically, the current direction is used to restore steps f and x f As input, the estimated value of the current diffusion step is obtained through the sequence denoising model. Afterwards, use and x f Get x f-1 , and so on, continuously iterating in the direction where the diffusion step tends to 0, and finally obtaining x0 after reverse reduction.

[0050] It should be noted that the model completes the inference prediction process of the item by matching the last item of x0 obtained with the embedded representation of the candidate set with maximum similarity.

[0051] 2. Designing a sequence denoising model

[0052] This embodiment designs a sequence denoising model based on Transformer, expressed as f θ (·), the model structure diagram is shown in Figure 2 In this embodiment, the forward process of the sequence denoising model is summarized as follows:

[0053]

[0054] Specifically, see formula (2), f θ The input of (·) is the sequence code E0 after being affected by the noise target term. The output is the sequence prediction result for the next moment Used for subsequent iterative denoising and restoration.

[0055] The above-mentioned influence of the noise target item on the input sequence is shown in formula (3). Taking the item at time t in the sequence as an example, represents the forward diffusion and noise-added encoding of the target item at time t+1; emb step Indicates the encoding of the reconstruction step the target item is currently in; t Indicates the project code at time t The degree of influence of the noisy target item information is different from the standard Gaussian distribution sampling method in other existing technologies. In order to integrate the long and short time position information of the current project to explore the importance of projects at different times, and then reveal the implicit relationship between the time when the project interaction occurs and the restoration of the target item, this embodiment designs a PosFuse network to learn λ t ,Depend on emb stepAnd the introduced project location code emb position Decision; t The original item information of the sequence and the noisy target item information are integrated by bitwise multiplication.

[0056] This embodiment uses user sequence encoding as input and uses the B2I dynamic routing algorithm to output a set of capsule interest vectors that can represent the user's diverse intentions. After extracting multiple interest capsules, this embodiment uses target item embedding and capsule interest vectors to complete label-aware attention calculation. For details, refer to formula (5), where p is the exponential index, and finally obtains multiple interest representations

[0057]

[0058] In order to correct the recommendation results of the sequence coding model and thus alleviate the recommendation bias, this embodiment uses the output of the sequence coding model As the target item, label-aware attention calculation is performed and finally we get It means: the item encoding at time t+1 predicted by the sequence denoising model for the item t∈[1,T-1] in the behavior sequence, and the user's multi-interest representation obtained through multi-interest extraction.

[0059] In order to eliminate the user-sensitive feature bias in the extracted multi-interest representation, this embodiment designs a user personalization tower and a regularization loss to correct the user multi-interest representation of the target item.

[0060] The training of this embodiment is divided into two stages, and the debiasing work mainly acts on the first stage. The features of the user personalized tower in this stage adopt user sensitive features, such as gender and age.

[0061] First, the user code output by the first stage user personalization tower is expressed as Combining multiple user interest representations The inner product regularization operation shown in formula (6) is completed as the auxiliary loss of the first stage training.

[0062]

[0063] For the user multi-interest representation of the output results of the sequence denoising model, this embodiment uses the above regularization method to eliminate the user sensitive feature bias contained therein. Afterwards, the present invention uses a multi-interest weighted method to correct the results of the sequence denoising model to improve the fairness of the recommendation results. At the same time, by adding user information to model personalization, the final denoising restoration and recommended item encoding is obtained through formula (7):

[0064]

[0065] in, At the same time, it should be noted that in the second stage of training and model reasoning, Replace with

[0066] 3. Model Training and Inference

[0067] In order to achieve a balance between recommendation diversity and personalization, the present invention designs a two-stage training method, wherein the first stage is responsible for fairness modeling and sequence denoising modeling; the second stage is responsible for adding user personalized information combined with user non-sensitive features.

[0068] First, since the length of some user behavior sequences does not meet the maximum sequence length set by the model, the present invention uses the input sequence S of the sequence denoising model as u (Except the last item) is filled, and according to the sequence position of the item, the expected output O = [o1, o2, ..., o T-1 ] is defined as follows:

[0069]

[0070] In order to prevent overfitting and take into account the training efficiency, this embodiment uses S u Each item in (except the last item) is negatively sampled once, denoted as G = [g1, g2, ..., g T-1 ], where G and O do not have repeated items. In addition, the sequence recommendation task often uses the potential embedding vector of the item to perform inner product matching to design the loss function. Combined with the task characteristics, this embodiment uses the binary cross entropy loss as the main loss function for model training to perform vector inner product matching between items. It should be noted that the model will ignore o t =The loss value at None is as follows:

[0071]

[0072] In order to achieve fair modeling, the first stage of training adopts a collaborative training method, which is responsible for optimizing the sequence denoising model and the personalized tower based on user sensitive features. The loss function is the weighted sum of the main loss function and the regular loss function, as shown below:

[0073]

[0074] Among them, λ controls the importance of fairness modeling. In the second stage of training, the present invention completely gets rid of user sensitive feature information, abandons the personalized tower based on user sensitive features, and instead uses user non-sensitive features to model user personalization. It should be noted that the present invention stops updating the parameters of the sequence denoising model at this stage, and only trains the newly added user personalized tower through the main loss function to further improve the recommendation accuracy.

[0075] In the inference stage, the present invention needs to combine the output O of the sequence denoising model with the reverse iterative algorithm mentioned above to obtain the final predicted item sequence code R, and then use the estimated code r of the last item T and the item codes in the candidate set i Perform inner product matching and select the final recommended items by maximizing them. The formula is as follows:

[0076]

[0077] It should be noted that in the field of generative images, pure noise obtained by sampling from a Gaussian distribution is generally used for reverse image restoration in the inference stage. However, in the field of recommendation systems, in order to inject user personalization, the present invention replaces pure noise with valid item codes in the user's historical behavior sequence, performs average pooling operations, and completes the result of N times of noise addition.

[0078] Example 2

[0079] The model constructed by the method for constructing a generative sequential recommendation model for recommendation fairness in Example 1 can improve the diversity of recommendation results.

[0080] The implementation modes of the present invention are described in detail above in conjunction with the embodiments, but the present invention is not limited to the above-mentioned implementation modes. For ordinary technicians in this technical field, after knowing the contents recorded in the present invention, they can make several equivalent changes and substitutions thereto without departing from the principle of the present invention, and these equivalent changes and substitutions should also be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for constructing a generative sequence recommendation model for recommendation fairness, characterized in that: The following steps are involved: Training phase: Inject random noise into user interaction behaviors to obtain a corrupted interaction representation that follows a Gaussian distribution; De-noising is then done in the reverse restoration process, providing a mapping from the corrupted state to the original state; Inference stage: imitate the forward denoising process to denoise the first n-1 items and then iterate to generate items in the reverse reconstruction process; Among them, random noise is injected into the user interaction behavior, and the damaged interaction representation that obeys the Gaussian distribution is obtained as follows: First, the interaction sequence of user u The Transformer sequence encoder is used to encode Right now Where, Embedding∈R |I|*d is an embedding representation vector table with dimension d; In the forward diffusion process, the interaction sequence Each item except the first one is embedded in Diffusion forward; and E n Expressed as and represents the initial encoding state of the item interacted by user u at the t∈[2,T]th moment when no diffusion occurs. The forward diffusion formula is as follows: in, β i Sampling is performed in a linearly increasing manner; ε obeys the standard Gaussian distribution Interaction Sequence Each project in The final code is represents the noise-added encoding of the items interacted by user u at the t∈[2,T]th moment after f diffusion processes; In the reverse restoration process of the training phase, a sequence denoising model is designed based on Transformer, and a dynamic routing mechanism is used to mine the multi-interest representations in the user's historical interaction behavior, combined with user-side modeling for two-stage training; In the first stage, a regularized auxiliary loss is added to eliminate the bias caused by user-sensitive features in the multi-interest representation, and then the processed multi-interest representation is used to correct the results of the sequence denoising model; In the second stage, the user’s remaining features are used to train the user’s personalized tower; In the reverse reconstruction process of the inference phase, items are generated iteratively with the help of posterior Bayes based on the trained model; Among them: Using dynamic routing mechanism to mine multiple interest representations in user historical interaction behaviors: using target item embedding and capsule interest vector to complete the label-aware attention calculation process, using user sequence encoding As input, the B2I dynamic routing algorithm outputs a set of capsule interest vectors that can represent the user's diverse intentions. After extracting multiple interest capsules, the output of the sequence encoding model is used As the target item, label-aware attention calculation is performed and finally we get Specific reference formula (1.2): in, p is the index, and finally we get the multi-interest representation The results of using the processed multi-interest representation to correct the sequence denoising model are as follows: the user personalized tower and regularization loss are used to correct the user multi-interest representation of the target item: First, the user code output by the first stage user personalization tower is expressed as Combining multiple user interest representations Complete the inner product regularization operation as shown in formula (1.3) as the auxiliary loss of the first stage training; Afterwards, the result of the sequence denoising model is corrected by using the multi-interest weighting method, and the final denoising restoration is obtained through formula (1.4) and used for the recommended item encoding in, In the second stage of training, the user information is added to personalize the model, and the user tower features are replaced with user non-sensitive features for model training. Replace with The item code is generated in the form of formula (1.4); in addition, the inference stage also uses the formula of the second stage training stage to generate the item code.

2. The method for constructing a generative sequential recommendation model for recommendation fairness according to claim 1, characterized in that: The forward process of the sequence denoising model is: In formula (2.1), f θ The input of (·) is the sequence code E0 after being affected by the noise target term. The output is the sequence prediction result for the next moment Used for subsequent iterative denoising and restoration; Formula (2.2) is the influence of the noise target item on the input sequence. represents the forward diffusion and noise-added encoding of the target item at time t+1; emb step The encoding of the reconstruction step that the target item is currently in; λ t Indicates the project code at time t The degree of influence of the noise target item information; In formula (2.3), λ t It integrates the original item information of the sequence and the noise target item information by bitwise multiplication. position Encodes the location of the item being imported.

3. A model constructed by the method for constructing a generative sequential recommendation model oriented towards recommendation fairness as described in claim 1 or 2.

Images